Reproduce “String Seed of Thought: Prompting LLMs for Distribution-Faithful and Diverse Generation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Kou Misaki, Takuya Akiba

Sakana AI introduces String Seed of Thought (SSoT), a prompting method that improves Large Language Models' adherence to probabilistic instructions and enhances output diversity. The approach achieved up to a 99% reduction in JS divergence for probabilistic tasks and yielded higher Distinct scores on the NoveltyBench benchmark compared to baseline methods.

Key insights

  • LLMs can autonomously devise and execute sophisticated strategies for randomness extraction and manipulation from internally generated strings, adapting them to task complexity.
  • The quality of LLM-generated internal randomness (uniformity, complexity) improves significantly as the length of its reasoning trace (Chain-of-Thought) increases.
  • Providing an internal, task-agnostic random string as a 'seed' allows LLMs to overcome inherent biases in direct probabilistic instruction following and mitigate diversity collapse.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.